gravityocr-dev

GravityOCR region-level OCR dev set: 402 cropped regions (Korean text, tables and formulas, plus English text), each with its prompt and target. Shared with participants for their own testing and tuning.

from datasets import load_dataset
ds = load_dataset("trillionlabs/gravityocr-dev", split="dev")      # 402 rows

Gated: request access on this page first and log in (hf auth login) with the account that was approved.

Composition

source	text	table	formula	total
blog	125	10	27	162
screenshot	2	3	1	6
authored	–	–	34	34
synthetic	–	50	–	50
aihub-479	33	–	67	100
omnidocbench (en)	50	–	–	50
total	210	63	129	402

Tasks

Each row is one cropped region, its prompt, and the target the model should output.

task	prompt	target format
text	Text Recognition:	plain text, with inline math as $…$
table	Table Recognition:	HTML <table>, header rows as td, merged cells as rowspan/colspan
formula	Formula Recognition:	bare LaTeX, with no $ (\tag is ignored when scoring)

Scoring

Scoring follows OmniDocBench's own metric code:

• text: 1 − normalized edit distance, after removing spaces and punctuation; BLEU is a secondary metric.
• table: TEDS; TEDS-S is a secondary metric.
• formula: CDM.

Overall = mean(text 1−NED, table TEDS, formula CDM) × 100. Report the score per source as well: the handwritten (aihub-479) and English (omnidocbench) parts have their own difficulty.

Columns

id, image, prompt, target, task_type, kind, source, source_detail, lang, degradation, width, height, gt_status, domain, note, render, difficulty.

gt_status describes the target:

• empty: exact by construction, i.e. rendered from the target;
• human_verified: screenshots;
• aihub_label: AI Hub human labels, which have some noise;
• omnidocbench_gt: OmniDocBench annotations.